A complete, compact, and detailed rulebook for building perfect LoRA datasets (FLUX, SDXL, SD3, Z-Image, etc.)#

This is written as a clear checklist, designed for humans and AI agents to follow exactly.

✔️ If you want the LoRA to learn a feature → label it#

✔️ If you want the LoRA to ignore a feature → do NOT label it#

✔️ If a feature is already strongly known to the base model → use the base word#

✔️ If a feature is new/custom → use a new invented token#

That’s the entire logic.

Everything else is the implementation below.

✔️ Include:#

• unique features
• custom designs
• special props
• specific hairstyle
• unique clothing patterns
• signature accessories
• distinctive body shapes
• unique color palettes

❌ Do NOT include:#

• scene
• background
• lighting
• mood
• camera angle
• composition
• common words you don’t want to bind

Because including them will force the LoRA to “bake in” those features.

Why?#

Without them, the LoRA may “replace” unknown words with something random.

Rule of thumb:

• Keep common descriptors that stabilize identity
• Remove common descriptors that you do NOT want to overtrain

Example list:

1
1girl, silver hair, long hair, high ponytail, blunt bangs, blue eyes, fair skin, slender build

Format:#

1
<activation_token><feature_token>

Examples:

1
xayahayaka_ribbon
2
xayahayaka_uniform
3
xayahayaka_hairbow
4
xayahayaka_tassel

Why do this?#

Because real words like:

1
ribbon
2
bow
3
pink ornament

are too general and conflict with the base model.

This method creates perfect controllable modular features that can be reused on other characters.

Strong recommended template for each training image#

1
<activation_token>,
2
(common identity descriptors),
3
(unique invented tokens),
4
(common clothing descriptors),
5
(unique clothing details),
6
(optional fixed color palette)

Example (your dataset, optimized)#

1
xayahayaka,
2
1girl, silver hair, long hair, high ponytail, blunt bangs, blue eyes, fair skin, slender build,
3
xayah_ribbon, xayah_tassel,
4
white sailor blouse, navy pleated skirt, opaque white pantyhose,
5
navy-and-gold color palette, pink tassel ornaments

✔️ Use the same invented token everywhere#

(e.g., always xayah_ribbon, never xayah ribbon or xayah-ribbon)

✔️ Describe the same feature with identical words#

(e.g., always “navy pleated skirt”, never mix it with “blue skirt”)

✔️ Keep body type words consistent#

(e.g., always “slender build”, not “slim” → “petite” → “thin”)

✔️ Use lower case only#

(helps tokenizer stability)

❌ Background#

Example: “cherry blossoms”, “park”, “street”, “sunlight”, “bench”

❌ Composition#

“low angle”, “full body”, “close-up”, etc.

❌ Emotional cues#

“smiling”, “sad”, “blushing”, “crying”

Because including them “freezes” these features into the LoRA.

Make sure:#

✔ Activation token is unique ✔ All unique features use invented tokens ✔ Common identity features included (if useful) ✔ Same words used consistently ✔ No background ✔ No camera angle ✔ No emotional features ✔ Describes ONLY what you want the LoRA to learn